A Review of Fault Tolerance Techniques in Generative Multi-Agent Systems for Real-Time Applications

Abstract

Low-light image enhancement remains a significant challenge in real-world computer vision applications, particularly in environments where lighting conditions are unpredictable and training data in the form of paired low/normallight images is unavailable. Inconsistent illumination, sensor noise, and non-uniform exposure settings severely degrade image quality, adversely impacting both human perception and the performance of downstream vision tasks such as object detection, recognition, and tracking. Although transformer-based models have recently demonstrated promising results on benchmark datasets under controlled conditions, their generalization to naturally degraded real-world images is still limited. Most existing models rely on synthetically paired datasets, which do not adequately capture the complexity and variability found in real-world low-light scenes.This paper introduces a novel approach to adapt a Retinex-inspired transformer model, Retinexformer, for real-world low-light image enhancement using unpaired data. We propose a domain adaptation framework that leverages unsupervised reconstruction losses, perceptual feature alignment, and domain-invariant regularization to fine-tune the model on naturally captured images. Our training pipeline eliminates the need for paired supervision, making the model more scalable and deployable across diverse applications. Extensive experiments on public real-world datasets such as ExDark and See-in-the-Dark (SID) demonstrate that our adapted model achieves superior performance in terms of both perceptual quality and quantitative metrics (PSNR, SSIM, LPIPS) when compared to baseline and state-of-the-art enhancement methods. Moreover, the proposed strategy significantly improves robustness in challenging low-light conditions while maintaining computational efficiency. Our findings show the potential of unpaired, real-world adaptation of transformerbased architectures for practical low-light imaging tasks including night photography, surveillance, and mobile vision systems.